Previsão de demanda de cosméticos no varejo utilizando aprendizagem de máquina

Abstract

Data-driven decision making was facilitated due to the high availability of data and the greater processing power of computers. To assist in decision making, it is possible to extract information from data through Data Science. An example in which there is great applicability of this science in companies is demand forecasting within the Supply Chain Management area. Forecasting sales volume is not a trivial task and inaccuracies in this forecast can cause stock-outs or affect its management. In this study, sales forecasts will be made for two different sales channels using Machine Learning algorithms for a brand owned by a large company. This company is in the Cosmetics, Fragrances and Toiletries market, where Brazil is the fourth largest consumer market in the world. Data was used from the years 2018 to 2023 on sales in all Brazilian states. Forecasts were made for three different time horizons: short term (next period), medium term (approximately 3 months ahead) and long term (approximately 7 months ahead). The short term refers to the next cycle for the regression methods and the next week for the time series method, the medium term refers to 5 cycles ahead for the regression methods and 15 weeks ahead for the time series method and the long term refers to the forecast of 10 cycles ahead for the regression methods and 30 weeks ahead for the time series method. Therefore, the consistency of the Machine Learning models was also evaluated. The algorithms analyzed in this study were CatBoost, LightGBM, XGBoost and Prophet. Firstly, the aforementioned Gradient Boosting methods were compared in order to identify which of the three methods showed the greatest stability when predicting multiple horizons. XGBoost had the lowest forecast errors for the Store channel in all three horizons (10% for the short term, 2.12% for the medium term and 6.4% for the long term). For the Direct Sales channel, XGBoost didn’t have the lowest WAPE in all horizons, but it was more stable compared to CatBoost and LightGBM. Next, XGBoost was compared with a time series method, Prophet. Comparing the two models in different scenarios, it was concluded that Prophet showed more satisfactory results and more stability in forecasting multiple time horizons.